diversitree (version 0.10-1)

find.mle: Maximimum Likelihood Inference

Description

Find the maximum likelihood point of a model by nonlinear optimisation. find.mle is generic, and allows different default behaviour for different likelihood functions.

Usage

find.mle(func, x.init, method, ...)
# S3 method for fit.mle
coef(object, full=FALSE, extra=FALSE, ...)
# S3 method for fit.mle
logLik(object, ...)
# S3 method for fit.mle
anova(object, ..., sequential=FALSE)

Value

A list of class fit.mle, with at least the components

• par The estimated parameters.

• lnLik The log likelihood at the ML point.

• counts The number of function evaluations performed during the search.

• code Convergence code. See the documentation for the underlying optimisation method for meaning, but "0" is usually good.

• func The likelihood function used in the fit.

• method The optimisation method used.

Arguments

func

A likelihood function. This is assumed to return the log likelihood (see Details). The function must take a vector of parameters as the first argument.

x.init

Initial starting point for the optimisation.

method

Method to use for optimisation. May be one of "optim", "subplex", "nlminb", "nlm" (partial unambigious string is allowed).

...

For find.mle, additional arguments passed through to the methods, optimisation routines, or to the likelihood function func - see Details. For anova, this is one or more models to compare against the model object (either submodels or supermodels or the test is meaningless).

object

A fitted model, returned by find.mle.

full

When returning the coefficients for a constrained model, should be coefficients for the underlying constrained model be returned?

extra

When returning the coefficients for a constrained model, should dummy “extra” parameters be returned as well?

sequential

Should anova treat the models as a series of increasing complexity? Currently this is a little overzealous in checking and will refuse to work if the likelihood values are not strictly increasing.

Model comparison

The anova function carries out likelihood ratio tests. There are a few possible configurations.

First, the first fit provided could be the focal fit, and all other fits are either special cases of it (every additional model is nested within the focal model) or generalisations of it (the focal model is nested within every additional model).

Second, the models could be sequential series of fits (if sequential=TRUE), such that models (A, B, C, D) are to be compared A vs. B, B vs. C, C vs. D. The models can either be strictly increasing in parameters (A nested in B, B nested in C, ...) or strictly decreasing in parameters (D nested in C, C nested in B, ...).

In both cases, nestedness is checked. First, the "class" of the fitted object must match. Second, the argnames of the likelihood function of a sub model must all appear in the argnames of the parent model. There are some cases where this second condition may not be satisfied and yet the comparison is valid (e.g., comparing a time-varying model against a non time varying model, and some make.quasse fits). We attempt to detect this but it may fail on some valid comparisons and silently allow some invalid comparisons.

Author

Richard G. FitzJohn

Details

find.mle starts a search for the maximum likelihood (ML) parameters from a starting point x.init. x.init should be the correct length for func, so that func(x.init) returns a valid likelihood. However, if func is a constrained function (via constrain) and x.init is the correct length for the unconstrained function then an attempt will be made to guess a valid starting point. This will often do poorly and a warning will be given.

Different methods will be dispatched for different types of likelihood functions. Currently all models in diversitree are supported (bisse, geosse, mk2, mkn, bd, and yule). With the exception of the Yule pure-birth process, these methods just specify different default arguments for the underlying optimisation routines (the Yule model has an analytical solution, and no optimisation step is required). Generally, it will not be necessary to specify the method argument to find.mle as a sensible method is chosen during dispatch.

The ... argument may contain additional arguments for the function func. This includes things like condition.surv for conditioning on survival in BiSSE, birth-death, and Yule models. Specify this as

    find.mle(lik, x.init, condition.surv=TRUE)
  

(see the Examples).

Different method arguments take different arguments passed through ... to control their behaviour:

method="optim": Uses R's optim function for the optimisation. This allows access to a variety of general purpose optimisation algorithms. The method within optim can be chosen via the argument optim.method, which is set to "L-BFGS-B" by default (box constrained quasi-Newton optimisation). This should be suitable for most uses. See the method argument of optim for other possibilities. If "L-BFGS-B" is used, then upper and lower bounds may be specified by the arguments lower and upper. The argument control can be used to specify other control parameters for the algorithms - see optim for details. Most of the optim algorithms require finite values be returned at every evaluated point. This is often not possible (extreme values of parameters or particular combinations may have zero likelihood and therefore -Inf log-likelihood). To get around this, the argument fail.value can be used to specify a fallback value. By default this is set to func(x.init) - 1000, which should work reasonably well for most cases.

method="subplex": Uses the "subplex" algorithm (a variant of the downhill simplex/Nelder-Mead algorithm that uses Nelder-Mead on a sequence of subspaces). This algorithm generally requires more evaluations than optim-based optimisation, but does not require approximation of derivatives and seems to find the global optimum more reliably (though often less precisely). Additional arguments are control to control aspects of the search (see subplex for details). The argument fail.value can be used as in method="optim", but by default -Inf will be used on failure to evaluate, which is generally appropriate.

method="nlminb": Uses the function nlminb for optimisation, so that optimising a Mk2/Mkn likelihood function behaves as similarly as possible to ape's ace function. As for method="optim", lower and upper bounds on parameters may be specified via lower and upper. fail.value can be used to control behaviour on evaluation failure, but like method="subplex", -Inf is used which should work in most cases. Additional control parameters may be passed via control - see link{nlminb} for details. This function is not generally recommended for use.

method="nlm": Uses the function nlm for optimisation, so that optimising a birth-death likelihood function behaves as similarly as possible to ape's birthdeath function. Takes the same additional arguments as method="nlminb" (except that fail.value behaves as for method="optim"). Like method="nlminb", this is not recommended for general use.

code and logLik methods exist for fit.mle objects so that parameters and log-likelihoods may be extracted. This also allows use with AIC.

Simple model comparison by way of likelihood ratio tests can be performed with anova. See Examples for usage.

Examples

## Due to a change in sample() behaviour in newer R it is necessary to
## use an older algorithm to replicate the previous examples
if (getRversion() >= "3.6.0") {
  RNGkind(sample.kind = "Rounding")
}

pars <- c(0.1, 0.2, 0.03, 0.03, 0.01, 0.01)
set.seed(2)
phy <- tree.bisse(pars, max.t=60, x0=0)

## Here is the 203 species tree with the true character history coded.
## Red is state '1', which has twice the speciation rate of black (state
## '0').
h <- history.from.sim.discrete(phy, 0:1)
plot(h, phy, cex=.5, show.node.state=FALSE)

## Make a BiSSE likelihood function
lik <- make.bisse(phy, phy$tip.state)
lik(pars)

## This takes ~30s to run, so is not enabled by default
if (FALSE) {
## Fit the full six-parameter model
fit <- find.mle(lik, pars)
fit[1:2]

coef(fit)   # Named vector of six parameters
logLik(fit) # -659.93
AIC(fit)    # 1331.86

## find.mle works with constrained models (see constrain).  Here
## the two speciation rates are constrained to be the same as each
## other.
lik.l <- constrain(lik, lambda0 ~ lambda1)
fit.l <- find.mle(lik.l, pars[-2])
logLik(fit.l) # 663.41

## Compare the models with anova - this shows that the more
## complicated model with two separate speciation rates fits
## significantly better than the simpler model with equal rates
## (p=0.008).
anova(fit, equal.lambda=fit.l)

## You can return the parameters for the full six parameter model from
## the fitted five parameter model - this makes a good starting point
## for a ML search.
coef(fit.l, full=TRUE)
}